General-purpose models keep getting better at benchmarks and keep disappointing in the clinic. The gap is rarely raw capability, it's everything around the model: which tools it can call, what context it gets, how the clinical task is decomposed, and how you know whether the output is actually right.This session covers the two halves of building vertical AI for clinicians, and why they only work together.

 

 

Part one: Engineering the harness. What a clinical agentic harness looks like in practice: context management, task decomposition, workflows, tool design, retrieval over messy longitudinal patient data, guardrails, and the runtime that makes iteration possible. Where straightforward agent designs break on real clinical data, and what we changed.

 

Part two — training the model inside the harness. How we train against harness-shaped tasks, how clinical eval sets and ground truth get built with clinicians in the loop, and why a smaller model trained inside the harness can beat a stronger generic one at the same job.

 

We close with open discussion. Aimed at engineers and researchers who want the systems view of applied AI in a high-stakes, regulated domain and anyone curious how a clinical AI team actually works day to day. Thought leadership, not a product walkthrough: the specifics come from our own stack, but the takeaways are meant to transfer to any vertical.